Changelog

0.1.1

A state found twice, once at desktop and again at the mobile viewport, no longer shows up as two mobile states.
visualice export never carries the viewer's development data into an export.
The plugin, skill and explorer agent run this version of the CLI (npx -y visualice@0.1.1).

The details panel's Live page link can point at a hosted copy of a site instead of the address it was crawled from (site.liveBase in a run's data), as the docs site's live demo does.

Docs site and live demo: alexdevlabs.github.io/visualice, now the package's homepage.

Crawl a website, capture every page and UI state across devices, and browse the shots on a zoomable canvas.

visualice crawl <url> discovers routes (links, sitemaps, data-loaded links, hash-router routes) and UI states (modals, menus, drawers, accordions, tabs, hover dropdowns, form focus and validation), then captures them on every configured device. A second probe pass at the first mobile viewport finds hamburger navs.
Full pages are captured one screen at a time, as a user scrolls, so sticky headers, cookie bars and modal backdrops show as they really do.
A cached site model makes later runs fast and deterministic (--replay); --resume continues an interrupted run. Hand-written and Explore-mode states survive re-crawls.
visualice.routes.md is an editable site index: add - /path to seed a page, - skip: to exclude one, - [ ] state to request a state.
Template sampling (scope.samples), site-wide states kept once, automatic isolation for states that write to storage, robots.txt and crawl-delay, retries, and a guard that fails a shot that lands on the login page.
Login recipes with saved sessions, staleness checks and credential redaction.
Other engines (devices[].browser: Firefox, WebKit), element-level component shots and axe-core accessibility checks.

crawl --dry-run lists the pages, states and screenshot count a crawl would take, without capturing. --no-states crawls pages only and keeps the states already cached. --seed=<paths> adds entry pages, --site=<name> picks the site folder a crawl writes to, and --headed shows the browser windows.
Clearer names, with the old ones still accepted (and a warning): --target (was --base), --dir (was --out), --timeout-ms (was --timeout), serve --screenshot (was --shoot). In the config, screenshot.fullPage is now "screens" | "single" | false (replacing fullPage + screens), maxHeight limits both modes (default 25000 CSS px; maxScreens still works), thumbnail.midWidth is zoomWidth, and interactions.auto: false is the one way to turn states off (nestDepth: 0 still works).

As a plugin or skill, exploring is the default: /visualice <url> maps the site with a dry run, confirms the page types and shot count with the user, then runs one visualice-explorer agent per page type. Each reads its page with visualice probe, verifies candidate recipes through the same code capture uses, and writes what it verified to an explore file; visualice author merges them into the cache and crawl --replay captures everything. --quick skips exploring.
/visualice-explore widens an existing catalogue: a new section, or a focus like form errors or signed-in states.

visualice serve: a dashboard of sites, each with dated versions, and a canvas per run, grouped by route with a row per state and a column per device.
Screen, Continuous and Scroll view modes; device and state-kind filters; search; minimap; light and dark themes; a phone layout.
Click a shot to focus it, with a floating bar of its actions (Full screen, Live page, Copy link, Download, Full image) matching the details panel.
Full screen shows a state on every checked device, with device captions.
Failed captures render as placeholder tiles; the view lives in the URL.
Sites can be deleted from the dashboard (from this computer only).
visualice export <dir> writes the viewer and a run as a static folder.

An npm package (npx visualice) with visualice doctor as a self-healing preflight, and a Claude Code plugin with the visualice skill, the explorer agent and three commands. The skill runs the CLI pinned to its own version (npx -y visualice@0.1.0).
AGENTS.md points other agents (Codex, Cursor, Gemini CLI) at the skill, with a by-hand version of the explore flow.
Needs Node 22.18 or newer, which runs a visualice.config.ts natively; older Node gets a clear message instead of a stack trace. Works on Windows: doctor's repairs run without npx.
init points the config's $schema at a versioned unpkg URL, so editors find it for global and npx installs too.

Things worth knowing before someone reports them as bugs.

No annotations in the viewer. Shots can be downloaded and linked to, not drawn on or commented on.
Discovery runs in Chromium only. A Firefox or WebKit device replays the recipes Chromium recorded; a state that only exists in another engine isn't found. Firefox has no mobile emulation, so a mobile: true Firefox device is a narrow desktop one.
scope.hashRoutes: "always" will fork on anchors. It exists for routers that drop the slash (#about); on a documentation site it turns every heading link into a route. auto is the default for that reason.
The Windows integration leg has never actually run. It's wired and the code it exercises is OS-sensitive (symlinked latest, run directories, temp paths); the first CI run on it is the one to watch.
Unnamed region signatures are a DOM path plus a hash of what the region says. Two unnamed regions with the same opening text in the same spot still collide, and a region whose text changes between probes (a live timer) stops deduping against itself. Named regions (id, aria-label, aria-labelledby) are unaffected, which is most of them.
A state that closes on scroll is shot without a post-open lazy-load pass. The page is scrolled before the state is opened instead, so content that only starts loading once the state is open and the page is scrolled isn't woken.